NumsAI's Kosilo cited alongside NVIDIA Kumo-Tabular and TabPFN-3.5 at GTC Berlin
The AMW Read
A small Korean tabular-model lab gets keynote visibility in a niche foundation-model sub-category with NVIDIA's backing; it updates the player map, but there is no deal or revenue signal yet.
Named counterparties: Prior Labs
NumsAI's Kosilo cited alongside NVIDIA Kumo-Tabular and TabPFN-3.5 at GTC Berlin
NumsAI is a Korean startup founded by Seoul National University researchers and led by CEO Jae-min Yoo. The company says NVIDIA CEO Jensen Huang named its structured-data foundation model, Kosilo, during his NVIDIA GTC Berlin 2026 keynote on October 21. Huang mentioned Kosilo while explaining tabular foundation models (TFMs), next to NVIDIA's own Kumo-Tabular and Prior Labs' TabPFN-3.5. NumsAI launched Kosilo in September. At the time, it said the model ranked first in the overall, classification and regression categories of the TabArena benchmark. The company has published the code and pre-trained weights. Commercial use requires a separate license.
Tabular foundation models are pre-trained on many structured datasets so they can handle new prediction tasks without a model built for each one. Typical tasks include demand forecasting, churn and sales estimation. The pitch is that this replaces much of the feature engineering and retraining behind classic machine-learning pipelines, which still run most enterprise forecasting. The article reports that NVIDIA sees TFMs as the next AI software contest after language models and acquired a company in the space this year for 600 billion won (~$430M). If that is accurate, the category now has a platform sponsor. NumsAI's mention reflects what NumsAI says was said on stage. It is not a commercial partnership.
For builders, the main question is whether open weights plus a paid commercial license can win enterprise forecasting work. Kosilo would compete with NVIDIA's own model and with Prior Labs. A TabArena lead is easy to check but quick to lose, so investors should watch for paying deployments in retail and services forecasting. Visibility at a keynote is not evidence of revenue.
